LibDebug: Make sure to not single step the program twice

After hitting a breakpoint, we single step the program to execute the
instruction we breaked on and re-enable the breakpoint.
We also single step the program when the user of LibDebug returned a
DebugDecision::SingleStep.

Previously, if we hit a breakpoint and then were asked to to a
DebugDecision::SingleStep, we would single step twice.

This bug can actually crash programs, because it might cause us to
skip over a patched INT3 instruction in the second single-step.

Interestingely enough, this bug manifested as functrace crashing
certain programs: after hitting a breakpoint on a CALL instruction,
functrace single steps the program to see where the CALL jumps to
(yes, this can be optimized :D). functrace crashed when a CALL
instruction jumps to another CALL, because it inserts breakpoints on CALL
instructions, and so the INT3 in the 2nd CALL was skipped over, and we
executed garbage :).

This commit fixes this by making sure not to single-step twice.
This commit is contained in:
Itamar 2020-05-23 17:11:11 +03:00 committed by Andreas Kling
parent 2686957836
commit f9d62fd5e5
Notes: sideshowbarker 2024-07-19 06:11:25 +09:00
2 changed files with 22 additions and 2 deletions

View file

@ -225,6 +225,12 @@ int DebugSession::continue_debugee_and_wait(ContinueType type)
void* DebugSession::single_step()
{
// Single stepping works by setting the x86 TRAP flag bit in the eflags register.
// This flag causes the cpu to enter single-stepping mode, which causes
// Interupt 1 (debug interrupt) to be emitted after every instruction.
// To single step the program, we set the TRAP flag and continue the debugee.
// After the debugee has stopped, we clear the TRAP flag.
auto regs = get_registers();
constexpr u32 TRAP_FLAG = 0x100;
regs.eflags |= TRAP_FLAG;

View file

@ -173,13 +173,20 @@ void DebugSession::run(Callback callback)
}
if (current_breakpoint.has_value()) {
// We want to make the breakpoint transparrent to the user of the debugger
// We want to make the breakpoint transparrent to the user of the debugger.
// To achieive this, we perform two rollbacks:
// 1. Set regs.eip to point at the actual address of the instruction we breaked on.
// regs.eip currently points to one byte after the address of the original instruction,
// because the cpu has just executed the INT3 we patched into the instruction.
// 2. We restore the original first byte of the instruction,
// because it was patched with INT3.
regs.eip = reinterpret_cast<u32>(current_breakpoint.value().address);
set_registers(regs);
disable_breakpoint(current_breakpoint.value().address);
}
DebugBreakReason reason = (state == State::Syscall && !current_breakpoint.has_value()) ? DebugBreakReason::Syscall : DebugBreakReason::Breakpoint;
DebugDecision decision = callback(reason, regs);
if (reason == DebugBreakReason::Syscall) {
@ -194,10 +201,17 @@ void DebugSession::run(Callback callback)
state = State::Syscall;
}
bool did_single_step = false;
// Re-enable the breakpoint if it wasn't removed by the user
if (current_breakpoint.has_value() && m_breakpoints.contains(current_breakpoint.value().address)) {
// The current breakpoint was removed in order to make it transparrent to the user.
// We now want to re-enable it - the code execution flow could hit it again.
// To re-enable the breakpoint, we first perform a single step and execute the
// instruction of the breakpoint, and then redo the INT3 patch in its first byte.
auto stopped_address = single_step();
enable_breakpoint(current_breakpoint.value().address);
did_single_step = true;
// If there is another breakpoint after the current one,
// Then we are already on it (because of single_step)
auto breakpoint_at_next_instruction = m_breakpoints.get(stopped_address);
@ -215,7 +229,7 @@ void DebugSession::run(Callback callback)
ASSERT_NOT_REACHED(); // TODO: implement
}
if (state == State::SingleStep) {
if (state == State::SingleStep && !did_single_step) {
single_step();
}
}